Uses old Rust 2015
|0.1.0||Feb 16, 2021|
#440 in Cryptography
12,167 downloads per month
Used in 8 crates (3 directly)
This is a thin fork of the
curve25519-dalek project, authored by
Isis Agora Lovecruft and Henry de Valence, in order to expose a formally
verified backed end supplied by the
fiat-crypto project, where
primitive curve operations are extracted from Coq proofs of arithmetic correctness.
A pure-Rust implementation of group operations on Ristretto and Curve25519.
curve25519-dalek is a library providing group operations on the Edwards and
Montgomery forms of Curve25519, and on the prime-order Ristretto group.
curve25519-dalek is not intended to provide implementations of any particular
crypto protocol. Rather, implementations of those protocols (such as
ed25519-dalek) should use
curve25519-dalek as a library.
curve25519-dalek is intended to provide a clean and safe mid-level API for use
implementing a wide range of ECC-based crypto protocols, such as key agreement,
signatures, anonymous credentials, rangeproofs, and zero-knowledge proof
curve25519-dalek implements Ristretto, which constructs a
prime-order group from a non-prime-order Edwards curve. This provides the
speed and safety benefits of Edwards curve arithmetic, without the pitfalls of
cofactor-related abstraction mismatches.
curve25519-dalek-fiat, add the following to the dependencies section of
curve25519-dalek-fiat = "0.1.0"
CHANGELOG.md for more details.
nightly feature enables features available only when using a Rust nightly
compiler. In particular, it is required for rendering documentation and for
the SIMD backends.
Curve arithmetic is implemented using one of the following backends:
fiat_u64_backendto use a verified u64 backend supplied by the
u32backend using serial formulas and
u64backend using serial formulas and
avx2backend using parallel formulas and
avx2instructions (sets speed records);
ifmabackend using parallel formulas and
ifmainstructions (sets speed records);
By default the
u64 backend is selected. To select a specific backend, use:
cargo build --no-default-features --features "std fiat_u64_backend" cargo build --no-default-features --features "std u32_backend" cargo build --no-default-features --features "std u64_backend" # Requires nightly, RUSTFLAGS="-C target_feature=+avx2" to use avx2 cargo build --no-default-features --features "std simd_backend" # Requires nightly, RUSTFLAGS="-C target_feature=+avx512ifma" to use ifma cargo build --no-default-features --features "std simd_backend"
curve25519-dalek can either select a backend on behalf of their
users, or expose feature flags that control the
std feature is enabled by default, but it can be disabled for no-
--no-default-features. Note that this requires explicitly
selecting an arithmetic backend using one of the
If no backend is selected, compilation will fail.
curve25519-dalek types are designed to make illegal states
unrepresentable. For example, any instance of an
guaranteed to hold a point on the Edwards curve, and any instance of a
RistrettoPoint is guaranteed to hold a valid point in the Ristretto
All operations are implemented using constant-time logic (no
secret-dependent branches, no secret-dependent memory accesses),
unless specifically marked as being variable-time code.
We believe that our constant-time logic is lowered to constant-time
assembly, at least on
As an additional guard against possible future compiler optimizations,
subtle crate places an optimization barrier before every
conditional move or assignment. More details can be found in the
documentation for the
Some functionality (e.g., multiscalar multiplication or batch inversion) requires heap allocation for temporary buffers. All heap-allocated buffers of potentially secret data are explicitly zeroed before release.
However, we do not attempt to zero stack data, for two reasons.
First, it's not possible to do so correctly: we don't have control
over stack allocations, so there's no way to know how much data to
wipe. Second, because
curve25519-dalek provides a mid-level API,
the correct place to start zeroing stack data is likely not at the
curve25519-dalek functions, but at the entrypoints of
functions in other crates.
The implementation is memory-safe, and contains no significant
unsafe code. The SIMD backend uses
unsafe internally to call SIMD
intrinsics. These are marked
unsafe only because invoking them on an
inappropriate CPU would cause
SIGILL, but the entire backend is only
compiled with appropriate
target_features, so this cannot occur.
Benchmarks are run using
cargo bench --no-default-features --features "std fiat_u64_backend" cargo bench --no-default-features --features "std u32_backend" cargo bench --no-default-features --features "std u64_backend" # Uses avx2 or ifma only if compiled for an appropriate target. export RUSTFLAGS="-C target_cpu=native" cargo bench --no-default-features --features "std simd_backend"
Performance is a secondary goal behind correctness, safety, and
clarity, but we aim to be competitive with other implementations.
fiat_u64 backend incurs a 8-15% slowdown compared to
u64 backend, depending on the EdDSA operation.
|Ed25519 batch signature verification/128||1.10 3.0±0.01ms||1.00 2.7±0.01ms|
|Ed25519 batch signature verification/16||1.09 411.7±1.28µs||1.00 377.8±0.92µs|
|Ed25519 batch signature verification/256||1.09 5.4±0.01ms||1.00 4.9±0.01ms|
|Ed25519 batch signature verification/32||1.08 779.3±4.87µs||1.00 723.9±3.21µs|
|Ed25519 batch signature verification/4||1.09 137.9±0.75µs||1.00 127.0±0.30µs|
|Ed25519 batch signature verification/64||1.15 1590.2±44.34µs||1.00 1385.2±6.80µs|
|Ed25519 batch signature verification/8||1.09 229.0±0.92µs||1.00 210.2±0.63µs|
|Ed25519 batch signature verification/96||1.11 2.4±0.08ms||1.00 2.2±0.01ms|
|Ed25519 keypair generation||1.07 17.9±0.07µs||1.00 16.7±0.09µs|
|Ed25519 signature verification||1.11 51.1±0.26µs||1.00 46.1±0.29µs|
|Ed25519 signing||1.05 18.9±0.09µs||1.00 18.0±0.07µs|
|Ed25519 signing w/ an expanded secret key||1.11 18.6±0.13µs||1.00 16.8±0.13µs|
|Ed25519 strict signature verification||1.06 53.0±0.33µs||1.00 50.0±0.15µs|